Abstract by Daniel Mortenson
Physics and Astronomy
Mapping the Ocean with Machine Learning
Source localization has been an important problem in underwater acoustics for decades. Traditional acoustic data analysis methods use models to simulate the ocean and then attempt to match the modeled data with the real data. This process is difficult, costly, and hard to generalize. By instead feeding raw spectrograms into a machine algorithm, our predictions will be more efficient, and likely more accurate than traditional methods. Additionally, data-driven techniques easily discern characteristics that are very difficult to discover otherwise, such as the type of seabed a sonar signal refracted off of.
Using data simulating shallow ocean off the coast of New England, our convolutional neural network algorithm predicts distance, ship speed, and seabed floor type at more than 95% accuracy after running for less than half an hour. Once run, the algorithm creates a model that can be loaded onto autonomous submarines, allowing for a more efficient way to map the ocean.